A0659
Title: Modelling extreme joint dependence using Bayesian nonparametric copulas
Authors: Concepcion Ausin - Universidad Carlos III de Madrid (Spain) [presenting]
Maria Kalli - Kings College London (United Kingdom)
Abstract: Modelling the joint distribution of financial variables is not an easy task. A flexible approach is needed in order to capture both the tail and central dependence structure as well as the slight asymmetry of the distribution. Copulas are useful in capturing the joint dependence of multiple random variables. However, parametric copulas are not flexible enough to capture both tail and central dependence, especially in periods of financial crises when tail dependence is higher and usually more important than overall correlations. From the Bayesian point of view, only a few nonparametric copula models have been proposed, and none of them adequately account for tail dependencies. A new Bayesian nonparametric copula is proposed, which can be viewed as a multivariate histogram smoothing with a non-equally spaced infinite number of bins. A prior is imposed on the breakpoints and on the volume of the bins. It is shown how to express the model as an infinite mixture of beta distributions using the stick-breaking representation of the Dirichlet process. Inference and prediction are made using a Gibbs sampler. The procedure is illustrated using simulated and real data based on multivariate financial time series.
